Some MACs require the PHY receive clock to be running to complete setup
actions. This may fail if the PHY has negotiated EEE, the MAC supports
receive clock stop, and the link has entered LPI state. Provide a pair
of APIs that MAC drivers can use to temporarily block the PHY disabling
the receive clock.
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Link: https://patch.msgid.link/E1tvO6k-008Vjt-MZ@rmk-PC.armlinux.org.uk
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The Synopsys Designware GMAC core databook requires all clocks to be
active in order to complete software reset, which we perform during
resume.
However, IEEE 802.3 allows a PHY to stop its clocks when placed in
low-power mode, which happens when the system is suspended and WoL
is not enabled.
As an attempt to work around this, commit 36d18b5664 ("net: stmmac:
start phylink instance before stmmac_hw_setup()") started phylink
early, but this has the side effect that the mac_link_up() method may
be called before or during the initialisation of GMAC hardware.
We also have the socfpga glue driver directly calling phy_resume()
also as an attempt to work around this.
In a previous commit, phylink_prepare_resume() has been introduced
to give MAC drivers a way to ensure that the PHY is resumed prior to
their initialisation of their MAC hardware. This commit adds the call,
and moves the phylink_resume() call back to where it should be before
the aforementioned commit.
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Link: https://patch.msgid.link/E1tvO6a-008Vjh-FG@rmk-PC.armlinux.org.uk
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When the system is suspended, the PHY may be placed in low-power mode
by setting the BMCR 0.11 Power down bit. IEEE 802.3 states that the
behaviour of the PHY in this state is implementation specific, and
the PHY is not required to meet the RX_CLK and TX_CLK requirements.
Essentially, this means that a PHY may stop the clocks that it is
generating while in power down state.
However, MACs exist which require the clocks from the PHY to be running
in order to properly resume. phylink_prepare_resume() provides them
with a way to clear the Power down bit early.
Note, however, that IEEE 802.3 gives PHYs up to 500ms grace before the
transmit and receive clocks meet the requirements after clearing the
power down bit.
Add a resume preparation function, which will ensure that the receive
clock from the PHY is appropriately configured while resuming.
Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Link: https://patch.msgid.link/E1tvO6V-008Vjb-AP@rmk-PC.armlinux.org.uk
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Commit 14a1968074 ("net: reorganize IP MIB values") changed
MIB values to group hot fields together.
Since then 5 new fields have been added without caring about
data locality.
This patch moves IPSTATS_MIB_OUTPKTS, IPSTATS_MIB_NOECTPKTS,
IPSTATS_MIB_ECT1PKTS, IPSTATS_MIB_ECT0PKTS, IPSTATS_MIB_CEPKTS
to the hot portion of per-cpu data.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://patch.msgid.link/20250320101434.3174412-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Kuniyuki Iwashima says:
====================
nexthop: Convert RTM_{NEW,DEL}NEXTHOP to per-netns RTNL.
Patch 1 - 5 move some validation for RTM_NEWNEXTHOP so that it can be
called without RTNL.
Patch 6 & 7 converts RTM_NEWNEXTHOP and RTM_DELNEXTHOP to per-netns RTNL.
Note that RTM_GETNEXTHOP and RTM_GETNEXTHOPBUCKET are not touched in
this series.
rtm_get_nexthop() can be easily converted to RCU, but rtm_dump_nexthop()
needs more work due to the left-to-right rbtree walk, which looks prone
to node deletion and tree rotation without a retry mechanism.
v1: https://lore.kernel.org/netdev/20250318233240.53946-1-kuniyu@amazon.com/
====================
Link: https://patch.msgid.link/20250319230743.65267-1-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
NHA_OIF needs to look up a device by __dev_get_by_index(),
which requires RTNL.
Let's move NHA_OIF validation to rtm_to_nh_config_rtnl().
Note that the proceeding checks made the original !cfg->nh_fdb
check redundant.
NHA_FDB is set -> NHA_OIF cannot be set
NHA_FDB is set but false -> NHA_OIF must be set
NHA_FDB is not set -> NHA_OIF must be set
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://patch.msgid.link/20250319230743.65267-4-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We will push RTNL down to rtm_new_nexthop(), and then we
want to move non-RTNL operations out of the scope.
nh_check_attr_group() validates NHA_GROUP attributes, and
nexthop_find_by_id() and some validation requires RTNL.
Let's factorise such parts as nh_check_attr_group_rtnl()
and call it from rtm_to_nh_config_rtnl().
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://patch.msgid.link/20250319230743.65267-3-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
ip6_rcv_core() is using:
__IP6_ADD_STATS(net, idev,
IPSTATS_MIB_NOECTPKTS +
(ipv6_get_dsfield(hdr) & INET_ECN_MASK),
max_t(unsigned short, 1, skb_shinfo(skb)->gso_segs));
This is currently evaluating both expressions twice.
Fix _DEVADD() and _DEVUPD() macros to evaluate their arguments once.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20250319212516.2385451-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
For CT action with commit argument, it's usually followed by the
forward action, either to the output netdev or next chain. The default
behavior for software is to drop by setting action attribute to
TC_ACT_SHOT instead of TC_ACT_PIPE if it's the last action. But driver
can't handle it, so block the offload for such case.
Signed-off-by: Jianbo Liu <jianbol@nvidia.com>
Reviewed-by: Roi Dayan <roid@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/1742392983-153050-6-git-send-email-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently, mlx5_is_reset_now_capable() checks whether the pci bridge is
accessible only on bridge hot plug capability check. If the pci bridge
is not accessible, reset now will fail regardless of bridge hotplug
capability. Move this check to function mlx5_is_reset_now_capable()
which, in such case, aborts the reset and does so in the request phase
instead of the reset now phase.
Signed-off-by: Aya Levin <ayal@nvidia.com>
Signed-off-by: Moshe Shemesh <moshe@nvidia.com>
Signed-off-by: Amir Tzin <amirtz@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/1742392983-153050-3-git-send-email-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
For simplicity reasons, the driver avoids crossing work queue fragment
boundaries within the same TX WQE (Work-Queue Element). Until today, as
the number of packets in a TX MPWQE (Multi-Packet WQE) descriptor is not
known in advance, the driver pre-prepared contiguous memory for the
largest possible WQE. For this, when getting too close to the fragment
edge, having no room for the largest WQE possible, the driver was
filling the fragment remainder with NOP descriptors, aligning the next
descriptor to the beginning of the next fragment.
Generating and handling these NOPs wastes resources, like: CPU cycles,
work-queue entries fetched to the device, and PCI bandwidth.
In this patch, we replace this NOPs filling mechanism in the TX MPWQE
flow. Instead, we utilize the remaining entries of the fragment with a
TX MPWQE. If this room turns out to be too small, we simply open an
additional descriptor starting at the beginning of the next fragment.
Performance benchmark:
uperf test, single server against 3 clients.
TCP multi-stream, bidir, traffic profile "2x350B read, 1400B write".
Bottleneck is in inbound PCI bandwidth (device POV).
+---------------+------------+------------+--------+
| | Before | After | |
+---------------+------------+------------+--------+
| BW | 117.4 Gbps | 121.1 Gbps | +3.1% |
+---------------+------------+------------+--------+
| tx_packets | 15 M/sec | 15.5 M/sec | +3.3% |
+---------------+------------+------------+--------+
| tx_nops | 3 M/sec | 0 | -100% |
+---------------+------------+------------+--------+
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Reviewed-by: Dragos Tatulea <dtatulea@nvidia.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/1742391746-118647-1-git-send-email-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Executing dql_reset after setting a non-zero value for limit_min can
lead to an unreasonable situation where dql->limit is less than
dql->limit_min.
For instance, after setting
/sys/class/net/eth*/queues/tx-0/byte_queue_limits/limit_min,
an ifconfig down/up operation might cause the ethernet driver to call
netdev_tx_reset_queue, which in turn invokes dql_reset.
In this case, dql->limit is reset to 0 while dql->limit_min remains
non-zero value, which is unexpected. The limit should always be
greater than or equal to limit_min.
Signed-off-by: Jing Su <jingsusu@didiglobal.com>
Link: https://patch.msgid.link/Z9qHD1s/NEuQBdgH@pilot-ThinkCentre-M930t-N000
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently, tcp_ao tests have two timeouts: TEST_RETRANSMIT_SEC and
TEST_TIMEOUT_SEC [by default 1 and 5 seconds]. The first one,
TEST_RETRANSMIT_SEC is used for operations that are expected to succeed
in order for a test to pass. It is usually not consumed and exists only
to avoid indefinite test run if the operation didn't complete.
The second one, TEST_RETRANSMIT_SEC exists for the tests that checking
operations, that are expected to fail/timeout. It is shorter as it is
fully consumed, with an expectation that if operation didn't succeed
during that period, it will timeout. And the related test that expects
the timeout is passing. The actual operation failure is then
cross-verified by other means like counters checks.
The issue with TEST_RETRANSMIT_SEC timeout is that 1 second is the exact
initial TCP timeout. So, in case the initial segment gets lost (quite
unlikely on local veth interface between two net namespaces, yet happens
in slow VMs), the retransmission never happens and as a result, the test
is not actually testing the functionality. Which in the end fails
counters checks.
As I want tcp_ao selftests to be fast and finishing in a reasonable
amount of time on manual run, I didn't consider increasing
TEST_RETRANSMIT_SEC.
Rather, initially, BPF_SOCK_OPS_TIMEOUT_INIT looked promising as a lever
to make the initial TCP timeout shorter. But as it's not a socket bpf
attached thing, but sock_ops (attaches to cgroups), the selftests would
have to use libbpf, which I wanted to avoid if not absolutely required.
Instead, use a mixed select() and counters polling mode with the longer
TEST_TIMEOUT_SEC timeout to detect running-away failed tests. It
actually not only allows losing segments and succeeding after
the previous TEST_RETRANSMIT_SEC timeout was consumed, but makes
the tests expecting timeout/failure pass faster.
The only test case taking longer (TEST_TIMEOUT_SEC) now is connect-deny
"wrong snd id", which checks for no key on SYN-ACK for which there is no
counter in the kernel (see tcp_make_synack()). Yet it can be speed up
by poking skpair from the trace event (see trace_tcp_ao_synack_no_key).
Fixes: ed9d09b309 ("selftests/net: Add a test for TCP-AO keys matching")
Reported-by: Jakub Kicinski <kuba@kernel.org>
Closes: https://lore.kernel.org/netdev/20241205070656.6ef344d7@kernel.org/
Signed-off-by: Dmitry Safonov <0x7f454c46@gmail.com>
Link: https://patch.msgid.link/20250319-tcp-ao-selftests-polling-v2-4-da48040153d1@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Rename __test_tcp_ao_counters_cmp() into test_assert_counters_ao() and
test_tcp_ao_key_counters_cmp() into test_assert_counters_key() as they
are asserts, rather than just compare functions.
Provide test_cmp_counters() helper, that's going to be used to compare
ao_info and netns counters as a stop condition for polling the sockets.
Signed-off-by: Dmitry Safonov <0x7f454c46@gmail.com>
Link: https://patch.msgid.link/20250319-tcp-ao-selftests-polling-v2-2-da48040153d1@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Before:
># 13145[lib/ftrace-tcp.c:427] trace event filter tcp_ao_key_not_found [2001:db8:1::1:-1 => 2001:db8:254::1:7010, L3index 0, flags: !FS!R!P!., keyid: 100, rnext: 100, maclen: -1, sne: -1] = 1
After:
># 13487[lib/ftrace-tcp.c:427] trace event filter tcp_ao_key_not_found [2001:db8:1::1:-1 => 2001:db8:254::1:7010, L3index 0, flags: S, keyid: 100, rnext: 100, maclen: -1, sne: -1] = 1
For the history, I think the initial format was to emphasize the absence
of flags as well as their presence (!R meant no RST flag). But looking
again, it's just unreadable and hard to understand.
Make it the standard/expected one.
Signed-off-by: Dmitry Safonov <0x7f454c46@gmail.com>
Link: https://patch.msgid.link/20250319-tcp-ao-selftests-polling-v2-1-da48040153d1@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Set metadata size building the skb from xdp_buff in cpsw/cpsw_new
drivers. ti cpsw and cpsw_new drivers set xdp headroom at least to
CPSW_HEADROOM_NA:
CPSW_HEADROOM_NA max(XDP_PACKET_HEADROOM, NET_SKB_PAD) + NET_IP_ALIGN
so the headroom is large enough to contain xdp_frame and xdp metadata.
Please note this patch is just compiled tested.
Reviewed-by: Michal Kubiak <michal.kubiak@intel.com>
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
Link: https://patch.msgid.link/20250318-mvneta-xdp-meta-v2-7-b6075778f61f@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Set metadata size building the skb from xdp_buff in mvpp2 driver
mvpp2 driver sets xdp headroom to:
MVPP2_MH_SIZE + MVPP2_SKB_HEADROOM
where
MVPP2_MH_SIZE 2
MVPP2_SKB_HEADROOM min(max(XDP_PACKET_HEADROOM, NET_SKB_PAD), 224)
so the headroom is large enough to contain xdp_frame and xdp metadata.
Please note this patch is just compiled tested.
Reviewed-by: Michal Kubiak <michal.kubiak@intel.com>
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
Link: https://patch.msgid.link/20250318-mvneta-xdp-meta-v2-2-b6075778f61f@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
There is an effort to achieve W=1 kernel builds without warnings.
As part of that effort Helge Deller highlighted the following warnings
in the tulip driver when compiling with W=1 and CONFIG_TULIP_MWI=n:
.../tulip_core.c: In function ‘tulip_init_one’:
.../tulip_core.c:1309:22: warning: variable ‘force_csr0’ set but not used
This patch addresses that problem using IS_ENABLED(). This approach has
the added benefit of reducing conditionally compiled code. And thus
increasing compile coverage. E.g. for allmodconfig builds which enable
CONFIG_TULIP_MWI.
Compile tested only.
No run-time effect intended.
Acked-by: Helge Deller <deller@gmx.de>
Signed-off-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20250318-tulip-w1-v3-1-a813fadd164d@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Kuniyuki Iwashima says:
====================
af_unix: Clean up headers.
AF_UNIX files include many unnecessary headers (netdevice.h and
rtnetlink.h, etc), and this series cleans them up.
Note that there are still some headers included indirectly and
modifying them triggers rebuild, which seems mostly inevitable. [0]
$ python3 include_graph.py net/unix/garbage.c linux/rtnetlink.h linux/netdevice.h
...
include/net/af_unix.h
| include/linux/net.h
| | include/linux/once.h
| | include/linux/sockptr.h
| | include/uapi/linux/net.h
| include/net/sock.h
| | include/linux/netdevice.h <---
...
| | include/net/dst.h
| | | include/linux/rtnetlink.h <---
[0]: https://gist.github.com/q2ven/9c5897f11a493145829029c0bfb364d0
====================
Link: https://patch.msgid.link/20250318034934.86708-1-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
net/unix/*.c include many unnecessary header files (rtnetlink.h,
netdevice.h, etc).
Let's clean them up.
af_unix.c:
+uapi/linux/sockios.h : Only exist under include/uapi
+uapi/linux/termios.h : Only exist under include/uapi
-linux/freezer.h : No longer use freezable_schedule_timeout()
-linux/in.h : No ipv4_is_XXX() etc
-linux/module.h : No longer support CONFIG_UNIX=m
-linux/netdevice.h : No dev used
-linux/rtnetlink.h : Not part of rtnetlink API
-linux/signal.h : signal_pending() is defined in sched/signal.h
-linux/stat.h : No struct stat used
-net/checksum.h : CHECKSUM_UNNECESSARY is defined in skbuff.h
diag.c:
+linux/dcache.h : struct dentry in sk_diag_dump_vfs()
+linux/user_namespace.h : struct user_namespace in sk_diag_dump_uid()
+uapi/linux/unix_diag.h : Only exist under include/uapi/
garbage.c:
+linux/list.h : struct unix_{vertex,edge}, etc
+linux/workqueue.h : DECLARE_WORK(unix_gc_work, ...)
-linux/file.h : No fget() etc
-linux/kernel.h : No cond_resched() etc
-linux/netdevice.h : No dev used
-linux/proc_fs.h : No procfs provided
-linux/string.h : No memcpy(), kmemdup(), etc
sysctl_net_unix.c:
+linux/string.h : kmemdup()
+net/net_namespace.h : struct net, net_eq()
-linux/mm.h : slab.h is enough
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250318034934.86708-5-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
include/net/af_unix.h indirectly includes some definitions for structs.
Let's include such headers explicitly.
linux/atomic.h : scm_stat.nr_fds
linux/net.h : unix_sock.peer_wq
linux/path.h : unix_sock.path
linux/spinlock.h : unix_sock.lock
linux/wait.h : unix_sock.peer_wake
uapi/linux/un.h : unix_address.name[]
linux/socket.h is removed as the structs there are not used directly,
and linux/un.h is clarified with uapi as un.h only exists under
include/uapi.
While at it, duplicate headers are removed from .c files.
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250318034934.86708-4-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
